Re: Updating Triton
Hi Kari,
On 2024-03-02 23:15, Kari Pahula wrote:
> I finally got MLIR and LLVM dependencies satisfied with libtriton.so.
Good to hear.
> Upstream's cmake was linking to both of those statically and when I
> changed it to dynamic linking the LLVM part stopped complaining as
> well. However, the way I did was heavily hacky as I pretty much ended
> with running sed on the cmake generated files. I couldn't find a
> cmake way to do it properly and I may have to ask upstream to do it.
> The style they're using is to is to download everything and use that
> copy with no care for sharing.
Yeah, this is an common pattern. This can often be addressed in cmake by
making the download conditional (USE_SYSTEM_LIBFOO), addressing the
linking type in the process. And this can easily be upstreamed.
Let us know when git has been updating, I'm sure someone will take a look.
> I still haven't got to testing this since solving the existing issues
> revealed the next blocker: dependency on pytorch with hip. I take
> that it's still in the works?
Yep, that's the white whale, so to speak.
See [1] for Cory's most recent report, in particular the attachment [2]
with the dependency graph for pytorch. There have been changes since then:
* rocalution is done
* rocdbgapi is RFS'ed [3], I think Petter is working on it
* hipify is currently RFS [4]
* I'm stalled on composable_kernel (infra work and package updates)
* you ITPed hipblaslt.
Best,
Christian
[1]: https://lists.debian.org/debian-ai/2023/12/msg00029.html
[2]: https://lists.debian.org/debian-ai/2023/12/pngpppSU99b6n.png
[3]: https://lists.debian.org/debian-ai/2024/02/msg00143.html
[4]: https://lists.debian.org/debian-ai/2024/02/msg00084.html
Reply to: